Fara

Active
GitHub Python MIT

Description

Microsoft's open-source browser and web task agent that uses large models to understand pages, plan actions, and complete real web workflows.

Key Features

  • 7B parameter agentic SLM designed for computer use with visual perception of webpages
  • Direct coordinate prediction for mouse and keyboard actions without accessibility trees
  • State-of-the-art results on WebVoyager, Online-M2W, DeepShop, and WebTailBench benchmarks
  • Compact size enabling on-device deployment with reduced latency and improved privacy
  • Efficient task completion averaging ~16 steps per task versus ~41 for comparable models
  • Novel synthetic data pipeline built on Magentic-One with 145K diverse training trajectories

Use Cases

💡 Automating web shopping tasks such as searching products and comparing prices across retailers
💡 Filling out forms, managing accounts, and completing online registrations
💡 Booking travel, movie tickets, and restaurant reservations through browser automation
💡 Searching for information and summarizing results from multiple web sources
💡 Researching job postings and real estate listings with multi-step navigation

Strengths & Limitations

Strengths

  • Actively maintained, recent updates
  • High community interest (6.1k stars)
  • Permissive open-source license (MIT)
  • Responsive to issues, low backlog

Quick Start

git clone https://github.com/microsoft/fara.git && cd fara && pip install -e . && playwright install && vllm serve "microsoft/Fara-7B" --port 5000

Related Projects

Stagehand

24.0k · TypeScript
Active A+

The SDK for browser agents by Browserbase. Provides act, extract, and observe primitives for AI agents to naturally browse and interact with web pages.

browser-agentsdkweb-automation +2
  • · Three-primitive API — act (execute actions), extract (get data), observe (inspect page)
  • · Hybrid natural language + code — freely switch between AI-driven and deterministic code control
  • · Auto-caching and self-healing — cache repeat actions to skip LLM inference, auto-fix on page changes

OpenBrowser

9.5k · TypeScript
Stale B

AI-powered autonomous web browsing framework that enables agents to click, type, navigate, and extract data like a human, with support for OpenAI, Anthropic, and Google models.

browser-agentweb-automationplaywright +2
  • · Natural language task execution — describe what you want and the agent navigates, clicks, types, and extracts data autonomously
  • · Multi-model support via Vercel AI SDK — swap between OpenAI, Anthropic, and Google with a single flag
  • · Interactive REPL for live browser sessions — drop into a browser prompt for debugging, prototyping, and exploration

AgenticSeek

26.9k · Python
Active A+

Fully local Manus AI alternative that autonomously browses the web, writes code, and interacts via voice, with no API costs

browser-agentcoding-agentlocal-ai +3
  • · Fully Local Execution — All data stays on-device with zero API costs and no cloud dependency
  • · Autonomous Web Browsing — Automatically searches, reads, extracts info, and fills web forms
  • · Coding Assistant — Supports writing and debugging in Python, C, Go, Java, and more